2016 Postmortem
Related: About this forumNate Cohn of NYT Upshot just casually confirmed difference in his and 538 methodology (Interesting!)
Nate Silver at 538 includes pretty much every single poll that comes out weighted by a number of factors including how "good" the pollster has been in the past. But everything gets included in the model. Everything (Excepting some known fake outfits). He has always maintained that even poor quality polls are useful for deducing trends.
Nate Cohn, Silver's replacement when he left the NYT for ESPN, has a model showing Clinton's probability of winning significantly higher than the one at 538. One of his followers asked how his model keeps showing Clinton leads in states versus aggregateors that have gone haywire thanks to firms like the Republican-owned, push-pulling Remington outfit flooding the market with narrative pushing polls.
Cohn's response shows a remarkably different philosophy in election modeling versus Silver's:
I basically only look at live interview polls and the few good online outfits https://t.co/PzHKkPhck3
https://twitter.com/Nate_Cohn/status/793821027755298816
In other words, he doesn't even bother with junk polls. Whereas, Silver prefers an "all data has value and the more the merrier" methodology.
And depending on which way you align, the narrative of the race looks different.
still_one
(92,216 posts)geek tragedy
(68,868 posts)Silver does a lot of clickbaity "The Clinton campaign is making a mistake" stuff as if Remington surveys are more insightful to Clinton's proprietary data machine
Cha
(297,275 posts)Adrahil
(13,340 posts)uses a "fat tails" model, meaning that it considers uncertainty more actively than a lot of other models. In other words, the 538 model considers the race to be more uncertain than the Upshot model. Both Nates are very good at what they do, IMO.
Let's not over-state this though.... both still consider Clinton a heavy favorite.
tallahasseedem
(6,716 posts)my nerves are SHOT!!!